How to backtest a strategy with Claude

Claude can backtest a strategy for you if it has a tool that runs backtests. ENSEMBLE provides one as a hosted MCP server, and connecting it takes about two minutes in Claude Desktop or one command in Claude Code. After that, describing a strategy in the conversation produces a real model with a public tear sheet, and asking Claude to change something produces a revised one. This guide covers the setup for both clients and then walks through a real session, including the places where the agent needed correcting.

Setup in Claude Desktop or claude.ai

1. Open Settings, then Connectors, then Add custom connector.
2. Paste https://api.ensemble.markets/mcp as the URL and click Add.
3. Claude opens the ENSEMBLE sign-in. Sign in or create an account. There are no keys to copy; the connection uses OAuth 2.1.
4. Start a new conversation. The ENSEMBLE tools are now available, and Claude will mention them when relevant.

The Free plan allows one custom connector; paid plans allow several. If you already have a connector configured and are on Free, you will need to remove it first.

Setup in Claude Code

From a terminal:

claude mcp add --transport http ensemble https://api.ensemble.markets/mcp

The first time a tool is called, Claude Code opens a browser for the ENSEMBLE sign-in. For headless or scripted use, create an ensemble_ API key in the dashboard and add it as a header instead:

claude mcp add --transport http ensemble https://api.ensemble.markets/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

Run /mcp inside Claude Code to confirm the server is connected and to see its tools.

A real session

What follows is a representative session in Claude Desktop, reconstructed around the real tool results for the dual momentum model published on this site. The user's messages are in bold. Tool calls are summarized in brackets; in Claude's interface the actual calls and results are visible if you expand them.

Backtest dual momentum on US and international stocks with a T-bill fallback.

I'll build that as a strategy. To make it testable I'm going to specify: SPY and EFA as the two equity funds, BIL as the cash position, a 12-month total-return lookback, and monthly evaluation at month end. Let me confirm those tickers exist and check their history first.

[calls list_assets for SPY, EFA, BIL]

All three are available. SPY and EFA both have data from 2001, so the tear sheet will open there; BIL launched in May 2007 and joins then. Building now.

[calls create_model with: "Dual momentum across SPY, EFA and BIL: hold whichever of SPY and EFA has the higher 12-month total return, and move to BIL when the leader's 12-month return is negative. Rebalance monthly."]

The build is running. I'll check on it.

[polls get_model three times over about 40 seconds]

Done. Here's what the model shows over May 2007 through the latest session:

• CAGR and volatility are on the tear sheet; the Sharpe ratio is about 0.5 and the maximum drawdown is around a third, versus roughly 55 percent for SPY over the same window.
• The run metadata shows about two rebalances per year on average, not twelve. That's expected: the rule is evaluated monthly but only trades when the signal changes.
• The QA note flags exactly that: the description says "monthly" and the backtest measured about 2 trades a year. That's the rule working as intended, but worth knowing if you expected more turnover.

The full tear sheet is at the link, including the equity curve against SPY, the drawdown chart, and the months it spent in each fund.

Two things in that answer are worth noticing. Claude confirmed the universe and the start date before building, which is the first habit of a careful backtest. And it read the run metadata and the QA finding rather than stopping at the headline metrics, which is how it caught the difference between "evaluated monthly" and "trades monthly."

What if we use a 6-month lookback instead?

I'll rebuild the model rather than build a new one, so the two versions stay linked.

[calls rebuild_model with the same description and "6-month" in place of "12-month"]

[polls get_model]

Here are the two tear sheets side by side. Compare the rebalances per year first: the shorter lookback reacts faster to turns and switches more often, and every switch is a cost. Then compare the drawdown charts and the compound returns over the same window. Whichever version looks better here, that's one sample, not a verdict; a different twenty years could rank them the other way.

The last sentence is the one to look for. An agent that says "the 12-month version is better" is overfitting on your behalf. One that says "over this window" is doing research. If Claude does not add the caveat, ask it what the difference would look like over a different window, and it will usually reason correctly about sample dependence.

Add a 60/40 core and hold 70 percent core, 30 percent dual momentum.

That's a portfolio: a fixed core plus a strategy sleeve. I'll build the core as its own strategy first, then compose them.

[calls create_model with "60% SPY and 40% TLT, rebalanced quarterly"]

[polls, then calls create_model with: "Hold 70% in m-… (the 60/40 core) and 30% in m-… (the dual momentum strategy), rebalanced quarterly."]

The portfolio's tear sheet is ready. Its holdings chart shows the equity weight stepping down whenever the dual momentum sleeve moves to T-bills, and the look-through allocation shows the actual asset weights at any date. Compare its drawdown chart with the two components' to see what the 30 percent sleeve changed.

This is the platform's composition model in one exchange: a model holds assets, other models, or both. Claude understood that "add a core" meant composition, not a more complicated single rule.

Prompts that get good results

The agent is only as precise as your request. These patterns produce testable models on the first try.

• Name the universe. "SPY, EFA and BIL" rather than "US and international stocks and cash." If you do not know the tickers, ask Claude to suggest liquid ETFs and confirm before building.
• Name the lookback and the cadence. "12-month return, evaluated at month end" rather than "recent performance."
• Say what happens when the signal is off. "Hold BIL" rather than leaving it implicit.
• Ask for one change at a time. "Change the lookback to 6 months" rather than "try some variations." The second produces a lot of models and no comparison.
• Ask Claude to read the code when a result surprises you. get_build returns the implementation (get_model names the current build_id), and Claude can explain what the rule actually did.

The common backtesting mistakes guide applies to agents as much as to people; an agent can overfit faster than you can.

Prompts to avoid

• "Find me the best strategy." This invites the agent to run many builds and report the winner, which is data snooping. Ask it to test a specific idea instead.
• "Optimize the parameters." Same problem. If you want to know how sensitive a result is to a parameter, ask for two or three conventional values and compare.
• "Should I invest in this?" The agent will decline, correctly. ENSEMBLE is research software, the backtests are hypothetical, and neither the platform nor Claude gives investment advice.

Reading what comes back

Whatever Claude summarizes, open the tear sheet. The metrics explained guide defines every figure on it. Check the start date against the funds' launch dates (get_model reports the window and which fund set it), check the universe against what you asked for, look at the drawdown chart before the return, and look at the trade count before believing the return survives costs. Claude is good at doing this for you and better at doing it when you ask.

Backtests are hypothetical, past performance does not guarantee future results, and ENSEMBLE is research software rather than an investment adviser.